:rocket: Lunar Lander with Genetic Algorithm

This project was highly inspired by Machine Learning Flappy Bird.

Lunar Lander is a single-player arcade game in the Lunar Lander subgenre. In the game, the player controls a lunar landing module as viewed from the side and attempts to land safely on the Moon.

For this project, I thought it would be a nice idea to combine this classic game with modern technology, and train these rockets to land by themselves.

How?

Inspired by Machine Learning Flappy Bird, I decided to combine the concept of Neuron Networks with Genetic Algorithm.

Neural Network

Each rocket is assigned with a neural network, with an input layer of 9 neurons, a hidden layer of 16 neurons, and an output layer of 4 neurons:

Inputs of the NN

9 input neurons:

• 8 inputs represent the surroundings of the rocket, telling the rocket the distances to the obstacles around it
• 1 input is the closest "landable" platform to the rocket

Outputs of the NN

The outputs of the NN comes in an array of 4 numbers between 0-1. Four of them each represents Thrust, Rotate Left, Rotate Right, and Do Nothing. If the value is over 0.5, the action is performed by the rocket.

Fitness

Each rocket has a fitness function that determines how successful a rocket is upon these criteria:

• Angle difference of landing/crashing platform
• Fuel used in rocket's lifetime
• Landing/crashing speed
• Distance of rocket to the closest landable platform

This is essentially the score of the rocket, used as a measure of how successful a certain rocket is at landing.

Genetic Algorithm

I mass spawn 30 rockets each iteration. In a certain time period, the rockets either crash, land or stay in the air.

After each iteration, I calculate the fitness of each and every rocket, and rank them from high to low. I keep the top rockets, and reproduce new rockets by crossing over and mutating their weights and biases out of the top units.

Results

After a few iterations, rockets start to land by themselves, and the average fitness slowly reaches a maximum value.

Resources

• Machine Learning Flappy Bird
• Coding Train's Evolution of Code
• My friend Baltazar's project
• Artistic inspiration from online lunar lander
• Synaptic NN Library